GildaVM: a Non-Blocking I/O Architecture for the Cog VM

Guillermo Polito (1) , Pablo Tesone (1) , Eliot Miranda (2) , David Simmons (3)

1 RMOD - Analyses and Languages Constructs for Object-Oriented Application Evolution
2 CDS - Cadence Design Systems
3 The Light Phone

The OpenSmalltalk virtual machine (VM) was historically designed as a single-threaded VM. All VM code including the Smalltalk interpreter, the garbage collector and the just-in-time compiler run in the same single native thread. While this VM provides concurrency through green threads, it cannot take advantage of multi-core processors. This architecture performs really well in practice until the VM accesses external resources such as e.g., FFI callouts, which block the single VM thread and prevent green threads to benefit from the processor. In this paper we present GildaVM, a multi-threaded VM architecture where one thread at a time executes the VM while allowing non-blocking I/O in parallel. The ownership of the VM is orchestrated by a Global Interpreter Lock (GIL) as in the standard implementations of Python and Ruby. However, within a single VM thread concurrency is still possible through green threads. We present a prototype implementation of this architecture running on top of the Stack flavour of the OpenSmalltalk VM. We finally evaluate several aspects of this architecture like FFI and thread-switch overhead. While current benchmarks show good results for long FFI calls, short FFI calls require more research to minimize the overhead of thread-switch.

Mots clés: virtual machines; multhreading; non-blocking I/O; foreign function calls

Domaines: Langage de programmation [cs.PL]

Guillermo Polito, Pablo Tesone, Eliot Miranda, David Simmons. GildaVM: a Non-Blocking I/O Architecture for the Cog VM. IWST19 - International Workshop on Smalltalk Technologies, Aug 2019, Cologne, Germany. ⟨hal-02379275⟩